Helix AI Workspace

An internal AI workspace that answers operational questions against live business systems with citations on every claim.

Generic assistants hallucinated against proprietary data and no team trusted the answers enough to act on them.

Interviews across four operational teams produced a labelled evaluation set of 900 real questions.

Hybrid retrieval with permission-aware filtering, answer grounding and a continuous evaluation harness gating releases.

1. 01Hybrid vector and keyword retrieval
2. 02Permission filter applied pre-ranking
3. 03Tool layer for live system reads
4. 04Evaluation harness in the release pipeline

Every prompt and retrieval change shipped only after passing the regression suite on the labelled question set.

1. Discovery

   Team interviews, labelled evaluation set.

2. Retrieval

   Hybrid index, permission filter, ranking.

3. Grounding

   Citation contract and refusal behaviour.

4. Rollout

   Team-by-team release gated on evals.

const context = await retrieve({
  query,
  filters: permissionsFor(user),
  strategy: "hybrid",
  topK: 12,
})

return answer({ query, context, requireCitations: true })

• Citation-backed answers
• Permission-aware retrieval
• Live system tool calls
• Answer feedback loop
• Team-level analytics

• Python
• TypeScript
• Supabase
• PostgreSQL
• Redis
• AWS

P95 answer

1.8s

Groundedness

96%

Eval pass

94%

Uptime

99.98%

Docs indexed

3.1k

Assisted search

+88%

CWV pass rate

100%

• Workspace
• Citation view
• Eval dashboard
• Admin console

Operational research time collapsed and the support team resolved tickets without escalation in most cases.